Equivariant Operator Discovery (EqOD)

• EqOD is a research framework that infers symmetry-compatible operators from data without pre-assuming the underlying group, enabling flexible discovery of latent representations.
• It employs multiple approaches—latent canonicalization, learned parameter-sharing, Koopman operator methods, sparse PDE identification, and algebraic classification—to capture equivariance across diverse applications.
• The paradigm offers improved out-of-distribution robustness and interpretability while posing challenges in operator identifiability, scaling to complex groups, and integration within deep architectures.

Equivariant Operator Discovery (EqOD) is the task of learning or enumerating operators whose input–output relationship respects symmetry, while inferring the relevant operator structure from data rather than assuming the symmetry group a priori. In the literature, the term covers several closely related programs: learning latent-space operators that canonicalize transformed observations, discovering equivariance through parameter-sharing schemes, identifying symmetry-respecting Koopman or neural operators, reducing PDE libraries by detected invariances, and constructing complete algebraic or geometric classes of equivariant operators (Yeh et al., 2022, Dinh et al., 20 Feb 2026, N'guessan et al., 12 May 2026). The name is also used specifically by “EqOD: Symmetry-Informed Stability Selection for PDE Identification,” a PDE-identification pipeline that combines symmetry detection, library reduction, sparse regression, and a residual-based fallback (N'guessan et al., 12 May 2026).

1. Core definitions and mathematical formulations

The common starting point is equivariance. A map $f$ is equivariant to a group $G$ acting on inputs via $g \cdot x$ and on outputs via a representation $\rho(g)$ if

$f(g \cdot x) = \rho(g) f(x).$

EqOD asks how such structure can be inferred or instantiated from data. In the latent-operator formulation, an encoder $\phi$ maps inputs to latent codes $z=\phi(x)$, and one learns latent operators $L_g$ such that

$L_g \phi(x) \approx \phi(g \cdot x),$

or, in canonicalization form,

$L_g^{-1}\phi(g \cdot x) \approx \phi(x).$

This places the discovery problem at the level of the representation $G$0 and the operator family $G$1 rather than the raw input architecture (Dinh et al., 20 Feb 2026).

A second formulation represents equivariance structurally through parameter sharing. For a linear map $G$2 under input–output permutations $G$3, equivariance is equivalent to the weight-tying constraint

$G$4

Learning the sharing pattern is therefore a way of learning the discrete group action itself. In “Equivariance Discovery by Learned Parameter-Sharing,” the sharing configuration is encoded by an assignment matrix $G$5 and free parameters $G$6, with tied parameter vector $G$7; the discovery problem is posed as a bilevel optimization over $G$8 (Yeh et al., 2022).

A third formulation appears in operator-theoretic and Koopman settings. There, EqOD learns a finite-dimensional surrogate operator $G$9 in the commutant of a representation $g \cdot x$0, enforcing

$g \cdot x$1

Under suitable assumptions on system equivariance, observable equivariance, and sampling-distribution invariance, the optimal EDMD matrix is itself equivariant, which permits group-convolutional parameterizations and block diagonalization by the generalized Fourier transform (Harder et al., 2024).

These formulations are not identical, but they share a common logic: symmetry is encoded as an operator relation, and the discovery problem consists in recovering the operator, the operator class, the canonical frame, or the symmetry-restricted hypothesis space from transformed data.

2. Principal methodological families

EqOD is not a single algorithmic template. The current literature contains several distinct families that differ in what is assumed known, what is learned, and where equivariance is enforced.

Family	Learned or fixed object	Representative paper
Latent canonicalization	$g \cdot x$2, $g \cdot x$3, or Lie-algebra coordinates for input/output transforms	(Dinh et al., 20 Feb 2026, Xu et al., 18 May 2026)
Parameter-sharing discovery	Sharing matrix $g \cdot x$4 and induced equivariant operator class	(Yeh et al., 2022)
Koopman / EDMD / OpInf	Equivariant $g \cdot x$5, stitched local operators, or reduced operators with symmetry coupling	(Nandanoori et al., 2021, Harder et al., 2024, Shuai et al., 24 Jul 2025)
Sparse PDE identification	Symmetry-reduced library and sparse coefficients	(N'guessan et al., 12 May 2026)
Algebraic or geometric classification	$g \cdot x$6-algebra operators, multigraph-indexed operator bases, isotropic kernels	(Hoyos et al., 19 May 2026, Ballerin et al., 16 May 2026, Shen et al., 2021)

One methodological axis concerns whether the group is known or discovered. Latent-operator approaches are described as learning latent operators “directly from data, using pairs or sets of transformed views, without requiring the group to be explicitly encoded in the network architecture,” although the fixed-operator variant still uses a generic cyclic shift template and the learned-operator variant uses a periodicity prior (Dinh et al., 20 Feb 2026). By contrast, PACE-FNO assumes known continuous symmetries of evolution equations on periodic domains and uses a learned Lie-algebra coordinate estimator only for the frame parameters, while the canonical operator itself is learned in the reduced frame (Xu et al., 18 May 2026).

Another axis concerns whether discovery is architectural, statistical, or algebraic. Learned parameter-sharing treats equivariance as a partition over parameters and evaluates recovery via the partition distance, which lower-bounds the discrepancy between equivariance groups (Yeh et al., 2022). Sparse PDE identification treats symmetry as a mechanism for pruning candidate libraries, for example by detecting Galilean invariance and removing terms ruled out by an exclusion theorem (N'guessan et al., 12 May 2026). Algebraic approaches treat equivariance as intrinsic to the multiplication law or the combinatorics of contractions: in the $g \cdot x$7 tensor algebra, equivariance is built into the product, while on constant-curvature spaces nonlinear polynomial differential operators are classified by equivalence classes of multigraphs (Hoyos et al., 19 May 2026, Ballerin et al., 16 May 2026).

A plausible implication is that “EqOD” is best understood as a research program organized around the discovery of symmetry-compatible operator structure, not as a single canonical architecture.

3. Latent operators, canonicalization, and robust recognition

The latent-operator approach in “Latent Equivariant Operators for Robust Object Recognition: Promise and Challenges” provides a minimal EqOD instantiation in which a linear encoder $g \cdot x$8 maps a flattened $g \cdot x$9 input to a $\rho(g)$0-dimensional latent with $\rho(g)$1, and a two-layer MLP classifier $\rho(g)$2 operates on canonicalized latents (Dinh et al., 20 Feb 2026). The paper studies discrete rotations, with 10 steps of $\rho(g)$3, and toroidal translations on a $\rho(g)$4 grid with stride 2 and order 14 per axis. The method learns $\rho(g)$5 jointly with either a pre-defined block-circulant shift operator or a learned $\rho(g)$6 operator with orthogonal initialization and a periodicity loss $\rho(g)$7, with $\rho(g)$8.

The training objective combines task supervision with canonicalization consistency. For transformed views $\rho(g)$9 and $f(g \cdot x) = \rho(g) f(x).$0, the model forms

$f(g \cdot x) = \rho(g) f(x).$1

and minimizes

$f(g \cdot x) = \rho(g) f(x).$2

together with cross-entropy on $f(g \cdot x) = \rho(g) f(x).$3. The total loss is

$f(g \cdot x) = \rho(g) f(x).$4

This makes canonicalization, rather than architectural group convolution, the central mechanism (Dinh et al., 20 Feb 2026).

The experiments use “noisy MNIST” with distractor backgrounds. Digits are thresholded, colored blue, then pasted over random black–white checkerboards; class 9 is excluded to avoid confusion with rotated 6. Training sees only rotations in $f(g \cdot x) = \rho(g) f(x).$5 and translations in $f(g \cdot x) = \rho(g) f(x).$6 pixels, while OOD consists of the remaining rotations and larger translations. Pose selection at inference uses a canonical reference bank and k-NN over candidate latent actions, with default main-plot settings $f(g \cdot x) = \rho(g) f(x).$7 and $f(g \cdot x) = \rho(g) f(x).$8 (Dinh et al., 20 Feb 2026).

The reported trends are sharply separated. The baseline using the same encoder and classifier but no operator peaks in-domain and collapses OOD: for rotation it reaches approximately $f(g \cdot x) = \rho(g) f(x).$9–$\phi$0 in-domain and drops to approximately $\phi$1–$\phi$2 at $\phi$3 and $\phi$4; for vertical translation it reaches approximately $\phi$5–$\phi$6 in-domain and drops to $\phi$7–$\phi$8 at far shifts. Operator models are flat across the full range. With ground-truth pose, both fixed and learned operators achieve approximately $\phi$9–$z=\phi(x)$0 across all rotations and translations; with automatic k-NN pose inference, rotation is approximately $z=\phi(x)$1–$z=\phi(x)$2 and translation approximately $z=\phi(x)$3–$z=\phi(x)$4 across the range. For compound translations, training uses only single-axis transforms, yet pre-defined and learned operators maintain high accuracy across most unseen $z=\phi(x)$5 combinations, and the learned operator is sometimes slightly better in corner regions (Dinh et al., 20 Feb 2026).

The paper positions this as an alternative to classic group-equivariant neural networks. G-CNNs and related architectures build equivariance by design and need the group a priori; latent EqOD instead learns usable operators in a learned latent space from transformed examples. The trade-off is explicit: latent EqOD offers flexibility and data-driven discovery of operators, but the paper emphasizes the lack of formal guarantees on OOD equivariance, open questions about operator placement in deep stacks, inference overhead from k-NN pose selection, and added difficulty for continuous, non-commutative, or 3D transformations (Dinh et al., 20 Feb 2026).

4. Dynamical systems, neural operators, and PDE identification

EqOD has developed particularly rapidly in dynamical systems and PDE learning. In the Koopman setting, “Data-Driven Operator Theoretic Methods for Phase Space Learning and Analysis” gives a symmetry-based transport theorem for local Koopman operators on invariant subspaces. If $z=\phi(x)$6 and the dictionary is consistent, then the local operators satisfy

$z=\phi(x)$7

and a global operator is assembled as a block-diagonal stitched operator

$z=\phi(x)$8

with stitched observables weighted by characteristic functions of the invariant subspaces. The same paper gives a residual criterion for when new data require a new dictionary and shows that the multiplicity of the unit eigenvalue increases as larger invariant subspaces are discovered (Nandanoori et al., 2021).

“Group-Convolutional Extended Dynamic Mode Decomposition” extends this line by showing that, under equivariance assumptions on the system, the observables, and the sampling distribution, the optimal EDMD matrix is equivariant and can be represented as a group convolution. In Fourier space, the operator decomposes into independent blocks indexed by irreducible representations, reducing both learning and spectral analysis to smaller problems. On the 2D Kuramoto–Sivashinsky PDE with $z=\phi(x)$9, GC-EDMD accurately learns dynamics using only 20 input-output pairs, while unconstrained EDMD fails; on the $L_g$0–$L_g$1 spiraling wave system with $L_g$2, it computes approximately $L_g$3 eigenvalues by solving 2304 independent $L_g$4 problems rather than one $L_g$5 problem (Harder et al., 2024).

A related shift-equivariant reduced-order setting appears in “Symmetry-reduced model reduction of shift-equivariant systems via operator inference.” There, the solution is rewritten in a traveling frame,

$L_g$6

so that

$L_g$7

After POD on aligned snapshots, the reduced dynamics acquire an explicit symmetry-coupling term,

$L_g$8

with a rational speed law

$L_g$9

For the Kuramoto–Sivashinsky equation, SR-OpInf with re-projection attains effectively zero training loss, matches intrusive SR-Galerkin, and yields relative snapshot errors of approximately $L_g \phi(x) \approx \phi(g \cdot x),$0 for $L_g \phi(x) \approx \phi(g \cdot x),$1 and approximately $L_g \phi(x) \approx \phi(g \cdot x),$2 for $L_g \phi(x) \approx \phi(g \cdot x),$3, while the non-reprojected model is unstable (Shuai et al., 24 Jul 2025).

PACE-FNO takes a canonicalization route closer in spirit to latent EqOD, but in PDE operator learning. It estimates a Lie-algebra coordinate $L_g \phi(x) \approx \phi(g \cdot x),$4, maps the input to a canonical frame, applies a standard Fourier Neural Operator $L_g \phi(x) \approx \phi(g \cdot x),$5, and restores the output to the target frame:

$L_g \phi(x) \approx \phi(g \cdot x),$6

The FNO core is unchanged; equivariance is enforced by the input and output transformations. On 1-D and 2-D Burgers, shallow-water, and Navier–Stokes equations on periodic domains, PACE-FNO matches the in-distribution accuracy of standard neural operators and reduces out-of-distribution relative error by up to $L_g \phi(x) \approx \phi(g \cdot x),$7 over FNO with symmetry augmentation under translations and Galilean shifts, with smaller gains for coupled rotation–translation shifts. In 1-D Burgers, the OOD relative error drops from $L_g \phi(x) \approx \phi(g \cdot x),$8 for FNO+Aug to $L_g \phi(x) \approx \phi(g \cdot x),$9 for PACE-FNO (one-shot) and $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$0 with test-time adaptation (Xu et al., 18 May 2026).

The 2026 PDE-identification method titled “EqOD: Symmetry-Informed Stability Selection for PDE Identification” uses the acronym in a narrower sense. It begins from a standard weak-form library

$L_g^{-1}\phi(g \cdot x) \approx \phi(x).$1

and combines two library-reduction mechanisms. If Galilean invariance is detected from trajectory data by a six-term weak-form structural test, the method uses a reduced library; otherwise it applies randomized LASSO stability selection. The sparse regression stage solves weak-form LASSO, chooses $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$2 by 5-fold cross-validation, applies adaptive thresholding and OLS debiasing, and then compares the reduced-library residual to the full-library residual through a fallback rule with $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$3 on the symmetry path and $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$4 on the stability path (N'guessan et al., 12 May 2026).

This PDE-identification EqOD comes with explicit benchmark results. On 8 PDEs at 4 noise levels, it attains $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$5 on Heat at $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$6 noise, where WF-LASSO obtains $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$7, official PySINDy 2.0 obtains $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$8, and the WSINDy reimplementation obtains $L_g^{-1}\phi(g \cdot x) \approx \phi(x).$9. Under the strict criterion that the mean F1 difference exceeds the larger of the two standard deviations, EqOD wins 7 of 32 cells, WF-LASSO wins none, and the remaining 25 cells are ties. External validation on WeakIdent and PINN-SR datasets gives $G$00 on all 5 clean benchmarks. At the same time, the paper is explicit that the Galilean library reduction is proved only under autonomy and library assumptions, and that formal guarantees for correlated PDE design matrices remain open (N'guessan et al., 12 May 2026).

5. Algebraic, representation-theoretic, and geometric viewpoints

A different EqOD strand treats equivariance as an algebraic object that can itself be scanned, decomposed, or classified. In “Group-Algebraic Tensors: Provably-optimal Equivariant Learning and Physical Symmetry Discovery,” the $G$01 tensor algebra makes equivariance intrinsic to multiplication. For tensors $G$02 and $G$03 with a group index, the product is

$G$04

and in the Fourier domain it decouples per irrep. The paper proves an Eckart–Young optimality theorem for the $G$05-SVD, shows that product groups factor through Kronecker products of generalized Fourier transforms, and uses a group scan over candidate $G$06 to score symmetry hypotheses. On full QM9, $G$07-SVD with ridge regression provides closed-form predictions at approximately $G$08–$G$09 fewer parameters than parameter-matched MLPs, and per-irrep decompositions recover selection-rule structure such as scalar properties being $G$10-dominated and dipole components being $G$11-dominated (Hoyos et al., 19 May 2026).

On geometric PDE operator spaces, “Equivariant nonlinear partial differential operators on constant curvature spaces” gives a classifying space for nonlinear differential operators on simply connected constant-curvature manifolds that are equivariant under the full isometry group. The operator class consists of nonlinear operators that can be written as polynomials in linear operators, and the paper shows that the classifying space is spanned by equivalence classes of multigraphs. For a multigraph $G$12,

$G$13

where vertices encode symmetric covariant derivatives and edges encode contractions by the metric. The map from graph classes to operators is surjective, and for graph classes with at most $G$14 edges and $G$15, the restriction is bijective. The same framework exposes dimension-dependent linear relations, such as the order-6 Hessian identities that vanish in $G$16 (Ballerin et al., 16 May 2026).

A more classical linear-operator perspective appears in “Rotation Equivariant Operators for Machine Learning on Scalar and Vector Fields.” There, linear, translation-invariant, rotation-equivariant operators on scalar, vector, and tensor fields are characterized as tensor-field convolutions with radially symmetric kernel fields. For vectors, isotropy forces the matrix kernel to take the form

$G$17

and differential operators such as gradient, divergence, curl, and Laplacian appear as distributional-kernel instances of the same theory. This viewpoint places EqOD in an explicitly interpretable function class: one discovers radial functions rather than unrestricted filters (Shen et al., 2021).

These algebraic and geometric formulations broaden EqOD beyond data-driven canonicalization. They frame symmetry discovery as basis discovery, per-irrep decomposition, group scanning, or operator-library generation. A plausible implication is that, in some domains, the primary EqOD object is not a neural architecture or latent code but a structured operator algebra.

6. Guarantees, trade-offs, and open questions

A recurrent issue in EqOD is the tension between flexibility and guarantees. Classic group-equivariant neural networks provide stronger theoretical guarantees but require the group a priori; latent EqOD explicitly trades those guarantees for flexibility and data-driven discovery of operators (Dinh et al., 20 Feb 2026). PACE-FNO, by contrast, restores a degree of analytical control by assuming known continuous symmetry groups and proving an OOD error decomposition

$G$18

which separates canonical-set approximation from alignment error (Xu et al., 18 May 2026).

Another central issue is identifiability. In learned parameter-sharing, the sharing matrix $G$19 is identifiable only up to permutation of cluster labels, and finite data can make spurious partitions locally optimal. The paper’s theoretical analysis is for Gaussian shared-mean estimation, where the mean-squared error decomposes into a bias term and a variance term depending on $G$20, and the finite-sample gap bound implies that, for strong sharing, relatively larger validation splits can be optimal—often much larger than the common 80/20 split (Yeh et al., 2022). This is a precise theoretical result, but it is tied to a restricted data model.

Several misconceptions are corrected by the existing literature. One is that EqOD necessarily removes the need for prior symmetry information. This is not uniformly true. The latent-operator formulation avoids explicit group-engineered layers but still uses cyclic templates or periodicity priors; PACE-FNO requires the continuous symmetry group and its field action; the constant-curvature classification presupposes the full isometry group; and GC-EDMD assumes an equivariant dictionary and group action (Dinh et al., 20 Feb 2026, Xu et al., 18 May 2026, Harder et al., 2024, Ballerin et al., 16 May 2026). Another misconception is that equivariance discovery is inherently neural. The literature includes bilevel partition learning, block-diagonal Koopman constructions, sparse regression over symmetry-pruned libraries, exact algebraic SVDs, and multigraph-indexed differential-operator bases (Yeh et al., 2022, Nandanoori et al., 2021, N'guessan et al., 12 May 2026, Hoyos et al., 19 May 2026).

Open problems are correspondingly diverse. Latent canonicalization faces open questions about operator placement in deep stacks, stability of learned operators, scaling to continuous, non-commutative, or 3D groups, and the computational cost of k-NN pose inference (Dinh et al., 20 Feb 2026). PDE-identification EqOD has a Galilean exclusion theorem under explicit assumptions, but formal guarantees for correlated weak-form PDE design matrices remain open (N'guessan et al., 12 May 2026). Constant-curvature operator classification is restricted to nonlinear operators that are polynomials in linear operators and to simply connected model spaces (Ballerin et al., 16 May 2026). The $G$21 algebra is finite-group based, so continuous symmetries are handled through finite subgroups or discretizations (Hoyos et al., 19 May 2026).

Taken together, these works suggest that EqOD is best regarded as a unifying perspective on symmetry-aware operator learning rather than a settled framework. Its central promise is to convert symmetry from a fixed architectural prior into a discoverable, testable, or classifiable property of operators. Its central challenge is that the meaning of “discovery” varies sharply across settings: discovering a latent operator, a parameter-sharing partition, a canonical frame, a sparse PDE library, a Koopman commutant element, or a multigraph basis are mathematically related but methodologically distinct enterprises.